In the same breath: how AI has changed the approach to lung CT scans | SBERMED AI
About company
Our solutions
Resource center

Contact us to find out more and start working with us

    * Required field

    SberMedAI LLC uses Cookies (files with information about past visits to the website) for service personalization and user convenience. To continue browsing the website, you need to enable cookies and accept our Cookies policy. You may disable cookies, but the website will become unavailable. Cookies policy


    In the same breath: how AI has changed the approach to lung CT scans

    June 22 2023


    Reading time 8 minutes

    Elena Sokolova, Chief of product office at SberMedAI, developer of artificial intelligence solutions in the field of medicine, talks about how artificial intelligence is changing the approach to lung computed tomography (CT).

    In May 2020, the Moscow government reported that artificial intelligence had processed 30,000 lung CT scans in Moscow in just two weeks. The technology has been actively used in hospitals and outpatient IT centers where the main stream of patients with COVID-19 is concentrated.

    At that time, there were already 114,431 patients in Russia. And after Moscow, other regions became interested in the use of artificial intelligence technologies in the interpretation of CT scans of the lungs.

    The Lung CT service was developed by Sber AI Lab and SberMedII to assist radiologists at the start of the pandemic. With the first release, we received a large number of requests from the regions and immediately started the piloting of the service. We understand what made the widespread use of AI in the processing of CT images during the pandemic so relevant, how this technology works today, and how exactly it is helping doctors.

    Computed tomography (CT) has long been a common diagnostic tool in medicine. But the advent of AI technologies has changed a lot, both in the way they conduct research and in the speed at which they interpret their results. The radiologist now has an assistant who can quickly process many images, highlight potentially dangerous areas on them, and immediately draw the doctor’s attention to the pathology. The AI ​​will calculate the rate of the lesion with high accuracy and generate a preliminary result which will then be verified by the doctor. Doctors needed this help, especially when hospitals and polyclinics began to receive hundreds and thousands of covid patients.

    At the very beginning of the pandemic, patients underwent CT scans to assess the dynamics of the disease on the first and third days after the start of treatment. The radiologists manually calculated the rate of the lesion and made a conclusion about the dynamics of the disease. Based on these data, the effectiveness of the treatment was evaluated and an effective treatment method was selected that could help more patients and save as many lives as possible.

    SberMedII understood these “pains” of doctors and therefore worked on the creation of technology that would simplify and speed up research processes, as well as rank patients by severity. It was important not to waste precious time and, above all, to analyze CT images of the most difficult patients, especially those who need to quickly monitor the dynamics of the disease.

    There was little time to create a service. In the background of the pandemic, the burden of radiologists increased day by day, they worked day and night. There were only two weeks to build the first version of the model, and the main challenge was finding doctors to flag the datasets collected, as all the radiologists were “in the field” and worked in several shifts.

    However, experienced radiologists from Moscow clinics took the time to participate in the project, marking the data and highlighting the affected areas in the images. The model was trained through studies involving more than 100,000 divisions in total, and the first pilots were commissioned in record time.

    It was a real test. The developers promptly gathered feedback, retrained models, and fixed technical issues. For example, during piloting in one of the regions, the model showed lower quality results. It turns out that clinics in the region are working with equipment whose images are not included in the initial training example (this problem is often encountered when creating medical AI services). The developers quickly created an additional dataset, retrained the model, and fixed the problem. As a result, the rollout of SberMedII developments in regional clinics and hospitals has reduced the time to obtain image analysis results by 20%. The developed model is continuously improved by being trained by an expert on new data.

    The process of the algorithm is divided into several stages: first, images are prepared in advance, then the image is segmented and pathological areas are automatically identified. In addition, AI recognizes objects, gives a quantitative assessment of identified pathological areas, and creates a preliminary conclusion with a description of the study. At the final stage, the result is confirmed by the radiologist of the clinic or, if necessary, by the doctor of the Medical Digital Diagnostic Center (MDDC), the SberMedII digital platform.

    Not only in Russia, but around the world, developers have developed technology for diagnosing lung diseases using artificial intelligence during the pandemic.

    For example, a group of scientists in the US tried to understand how well different AI models could detect COVID-19 symptoms from chest CT scans.

    The developers tested two classification models: a full 3D model examined the entire region of the lungs, transforming the image according to the specified dimensions, and a hybrid 3D model created an image based on several tomographic sections. A total of 2,724 scans of 2,617 patients, including those with confirmed coronavirus, were used to train and test the technology.

    The experiment showed that the full 3D model correctly identified the symptoms of COVID-19 in 87 of 109 patients and the hybrid model in only 74 of 109 patients. The scientists also noticed that the machine was getting better and worse at recognizing early pneumonia. diseases in the diagnosis of a progressive course.

    Chinese scientists immediately became interested in the problem. Neural networks CovNet have learned to distinguish between symptoms of coronavirus and signs of community-acquired pneumonia. The working principle of the three-dimensional model was presented in the journal Radiology: as input, the network takes parts of a CT image, then generates a function for the corresponding parts, combines the extracted elements with a function map, and finally estimates the probability of each pathology – pneumonia in COVID-19 , community-acquired pneumonia and non-pneumonic diseases. To visualize the affected areas, the scientists introduced heatmaps. The algorithm highlights suspicious areas in red for each predicted class, thus focusing the radiologist’s attention on them.

    It turned out that CovNet correctly detected covid pneumonia in the vast majority of cases. The sensitivity of the network in detecting COVID-19 was 90% and the specificity was 96%.

    Alongside the diagnosis of covid-19, scientists have not stopped work that has begun to identify signs of other diseases in CT scans using AI, the same lung cancer that kills about 1.6 million people worldwide each year. For example, a group of scientists at the Massachusetts Institute of Technology (MIT) developed a deep learning model for predicting future cancer risk with the loud name Sybil (similar to the ancient Greek sibyl wandering prophets). The model was trained on images collected during the national lung screening study conducted from 2002 to 2004. Before the team tested Sybil with no clear signs of cancer on CT scans, they had labeled hundreds of scans with open malignancy to teach the machine to adequately assess all risks. During the experiment, it became clear that the model saw potential risks even when people failed to detect a malignant tumor in the body.

    There were also studies in this direction in Russia. Therefore, in 2021, the developers of SberMedII trained the Lung CT algorithm to identify early signs of cancer. The machine marks possible pathology areas with colored clues, sorts the medical images by probability of pathology, and generates a preliminary result. If AI “sees” a 4 mm neoplasm, it highlights all nodules in the image, regardless of their size.

    The examination is then reviewed by the facility’s radiologist or sent to the MDDC physician for verification. If a person has a high risk of oncology, he is immediately sent to an oncology dispensary, where doctors decide on his additional examination and further referral.

    The enhanced Lung CT service was piloted in the Nizhny Novgorod region in April-May 2022. AI analyzed 5,121 studies there and found 184 possible cases of neoplasm. After confirmation by doctors, the results of 124 lung CT scans were sent to the Nizhny Novgorod Cancer Center.

    Then, in October 2022, a technology pilot was conducted at the Regional Clinical Hospital No. 3 in Chelyabinsk, where AI analyzed 261 studies and identified 68 suspected cancers. Thirteen cases of suspected malignancy were confirmed by doctors.

    Today, AI-enabled solutions have already demonstrated their effectiveness in analyzing CT scans of the lungs. The possibilities of such diagnostics are expanding every day, multi-purpose services are being created to simultaneously detect several pathologies with CT of the chest organs.

    AI detects any subtle pathology, reducing image interpretation time and increasing patient throughput while increasing clinical confidence in results. It usually takes a lot of time for a doctor to measure and calculate the lesion area, and the service does this automatically, quickly and accurately calculate the percentage of lung lesion, the doctor only needs to check and confirm the results. In the future, we may get to the point where all medical imaging studies will first be interpreted by AI algorithms and then verified by a doctor. This approach will reduce the burden of radiologists and the number of diagnostic errors, as the doctor usually focuses on finding a specific pathology, and the multi-target algorithm always acts in the same way and the human eye can see what it is. may not be able to distinguish by mistake.

    The author expresses his personal opinion, which may not coincide with the editors’ position.

    Source: Газета.Ru

    Image: Даша Зайцева/«Газета.Ru»

    Другие новости